A comprehensive deep learning approach for harvest ready sugarcane pixel classification in Punjab, Pakistan using Sentinel-2 multispectral imagery

Sugarcane is an important crop for the production of sugar and ethanol, and its area has increased significantly in recent decades in tropical and subtropical regions. Pakistan is among the top ten producers of sugarcane in the world. Up-to-date and accurate sugarcane maps are critical for monitoring sugarcane acreage, and production and assessing its social, economic, and environmental impacts. A huge amount of work has been published regarding crop monitoring and mapping using remote sensing techniques. This study proposes a deep learning-based framework for pixel-based classification of the sugarcane crop among other popular crops grown in Pakistan (e.g., rice, wheat, and corn) using Sentinel-2 multispectral imagery. The frame work includes selection of Sentinel products (Level-2A), preprocessing, spectral indices extraction, spectral feature compilation, labeling through spectral unmixing and harvest time sugarcane classification. The selection of Sentinel products for each crop field is based on the NDVI values. Different spectral (NDVI, NDWI, DVI, SAVI) and biophysical indices (LAI, FVC) are extracted from these sentinel products. Every pixel is compiled as a 2D feature map containing the time-series (ten-time stamps) evolution of each pixel across twelve spectral bands and six indices. The time-series multispectral feature maps are subjected to bilinear sampling to prepare them for input into different deep learning models. Moreover, the labeling of each pixel is done using linear spectral unmixing to assure the abundance of that relevant crop in each pixel. The data set contains samples from different districts of Pakistan and two combinations of a dataset are formed to check the robustness of the developed methodology i.e., training and testing from the same district and from separate districts. For the first combination of datasets, the F1 score for most of the classification models (Convnet, VGG16, ResNet-50, Inception v3 and LSTM) tested is high nearly about 0.99, and for the second set, LSTM outperformed other models with the F1 score of 0.9. The classified pixels are seamlessly integrated into the classification maps of the respective fields. The harvest time sugarcane classification yields encouraging results when compared to the ConvNext and shows high potential to classify sugarcane among other crops using few numbers of products and is capable of classifying sugarcane from other districts as well.